# Session 2 Handout: Summarizing for Real Readers

<p class="meta">FAST-02 Generative AI for Everyday Work · 7formations · Instructor: Wadood Chaudhary · All examples fictional</p>

## Why this session matters

"Summarize this" is the most common AI request at work, and the least specified. A good summary is written for a particular reader who must do something afterwards. A trustworthy summary has been checked against the source. This session covers both, using a 30-page safety manual from North Shore Builders (fictional, Beverly).

## Key concepts

### 1. Choose the summary type first

| Type | Reader needs | Shape | Example |
| --- | --- | --- | --- |
| Brief | Know what it is and why it matters | 3 to 5 lines | "The 2026 manual adds a heat rule and weekly ladder inspections." |
| Decision memo | Choose between options | Options, recommendation, open questions | "Adopt vendor A or B for harness inspections?" |
| Action list | Do things | Who, what, by when | "Foremen: restock first-aid kits by Friday." |

Name the type in the prompt. Without it you get a generic recap that serves nobody.

### 2. Write for the reader, then set the length

Name the reader (busy owner, new hire, customer, inspector) and what they must do after reading. Then set a hard limit: lines, bullets, or words. Ask for plain language and defined terms. Remember that **shorter summaries need more checking, not less**: to compress, the model must choose, merge, and round, and that is where numbers drift and rules flip.

### 3. Five-fact verification

The minimum check for any summary someone will rely on:

1. Pick five facts from the summary: numbers, dates, names, rules, deadlines.
2. Find each one in the source and mark the location.
3. Record **match**, **wrong**, or **missing** (stated in the summary but not found in the source).
4. Any wrong or missing fact: fix the summary and recheck.
5. Keep the log. It is your evidence that you checked.

Pick the facts that would cause the most harm if wrong: safety thresholds, deadlines, money.

### 4. Long files: limits and chunking

Every tool has a limit on how much it reads well. Even with very large context windows, details from the middle of long documents are more likely to be dropped or blended. The reliable approach:

- Split by section (or chapter, or date range).
- Summarize each part, asking "what did you leave out of this section?"
- Combine the section summaries for each reader.
- Check the combined summary against the document's headings: which sections are missing, and is that right for this reader?

Tools that hold documents as sources (NotebookLM, Copilot referencing a file, a Claude or ChatGPT project) help with retrieval and citations, but the coverage and five-fact checks still apply.

## Worked example: one manual, two readers

**Source facts (fictional NSB manual):** hard hats in all active work zones, including the yard; fall protection at 6 feet; ladders 3 feet above landing; incidents reported to supervisor within 1 hour; new heat rule (10-minute break per hour above 90°F heat index); ladder inspections now weekly.

**Supervisor brief (decision-oriented, 120 words):** leads with what changed from 2025 (heat rule, weekly ladder inspections), what to enforce this week, and open questions (who logs heat breaks?).

**New-hire list (plain, 90 words):** six exact rules: hard hat everywhere work is active including the yard; harness or guardrail at 6 feet; ladder 3 feet above the landing, never stand on the top step; tell the supervisor within 1 hour of any incident; water, shade, 10-minute break each hour when it's hot; ask if unsure.

**Five-fact check (new-hire list, first AI draft):**

| Fact | Source | Result | Fix |
| --- | --- | --- | --- |
| Fall protection at 10 ft | §4: 6 ft | Wrong | Changed to 6 ft |
| Ladder 3 ft above landing | §5 | Match | |
| Report by end of shift | §7: within 1 hour | Wrong | Changed to 1 hour |
| Hard hats on site | §2: all active work zones incl. yard | Partly | Tightened wording |
| Heat break 10 min/hour | §8 | Match | |

## Checklist: summary ready to send

- [ ] Summary type named (brief, decision memo, action list).
- [ ] Reader and their next action named.
- [ ] Hard length set and met.
- [ ] Long file chunked; coverage checked against headings.
- [ ] Five facts checked and logged; errors fixed and rechecked.
- [ ] Same facts across versions for different readers.

## Do and don't

| Do | Don't |
| --- | --- |
| Name the reader and what they must do | Ask for "a summary" and hope |
| Check the five highest-harm facts | Check only the first paragraph |
| Chunk long files and ask what was left out | Paste 30 pages and trust a one-paragraph answer |
| Keep facts identical across audiences | Let the new-hire version "simplify" a safety number |
| Log your checks | Rely on memory that you checked |

## Try this at work

1. Pick a long, non-sensitive document you or your team rely on.
2. Write two summary prompts for two readers. Run them.
3. Do a five-fact check on each and record the result.
4. Share the better one with a coworker and ask what they'd do after reading it. If they can't say, revise.

## Glossary

| Term | Meaning |
| --- | --- |
| Brief | Short "what and why it matters" summary |
| Decision memo | Summary built around options and a recommendation |
| Action list | Who does what by when |
| Five-fact check | Verifying five key facts against the source and logging results |
| Chunking | Splitting a long document into sections before summarizing |
| Coverage check | Comparing a summary against headings to find omissions |

## Key takeaways

- Type and reader first; length second.
- Shorter means more checking.
- Five facts, logged, every time.

*This manual and its rules are fictional training material and are not OSHA requirements. Check current federal and state rules for real worksites.*